CN101477801A - Method for detecting and eliminating pulse noise in digital audio signal - Google Patents

Abstract

The invention provides a method for detecting and removing pulse noise in a digital audio signal, which comprises the following steps: utilizing weighted difference vectors of adjacent signal samples in a certain neighbor in the signals to detect pulse number and determine detection thresholds of the noise pulse according to the estimation value of the change degree of the amplitude and frequency of the signal; utilizing the envelope based on the weighted difference vectors of the signal samples to evaluate the width of the noise pulse and establish an autoregressive model of the signal; according to the correlation of voice frequency signals, utilizing adjacent signal samples which are repaired or not contaminated to repair a signal section which is contaminated by the pulse noise and recover original signals. The invention has the advantages that a normal signal which is changed abruptly and the noise pulse can be separated, and the width of the pulse can be calculated; and meanwhile, by adopting the envelop which is formed by the weighted difference vectors of the signal, the position and the width of the noise pulse are detected, so that a single noise pulse, a plurality of contiguous noise pulses or a plurality of overlapping noise pulses can be processed.

Description

A kind of method that detects and eliminate impulsive noise in the digital audio and video signals

Technical field

The present invention relates to a kind of method that detects and eliminate impulsive noise in the digital audio and video signals, be used for the detection and the noise reduction of digital audio and video signals impulsive noise, belong to the detection and the noise-reduction method technical field of digital audio and video signals impulsive noise.

Background technology

The many forms with simulating signal of traditional audio data are kept in the media such as tape, disc.These data to be preserved for a long time, digital signal need be converted thereof into.In digitized process, owing to reasons such as aging, the damage of storage medium, defective, conversion equipments, have the noises that influence the sense of hearing in a large number in the digital audio and video signals after the conversion, reduce or eliminate these noises, the process of recovering original sound signal is called audio defeat.

Impulsive noise is also referred to as the Click noise, is one of overriding noise that influences in the sound signal people's sense of hearing.Its duration of single noise spike is generally less than 100 milliseconds.Because impulsive noise continues to grow, changes in amplitude is bigger, is difficult to come noise reduction with noise-reduction method commonly used as Fourier transform, wavelet de-noising, Hi-pass filter.Traditional impulse noise detection mainly adopts the upset wave filter to carry out, and this method is difficult to the unexpected signal that changes in the normal signal as sharp-pointed knock, is made a distinction with noise spike, very accurately the estimating noise width; Traditional impulsive noise restorative procedure adopts median filter more, and its shortcoming is to be difficult to obtain high-quality sound signal repairing effect.

Summary of the invention

The purpose of this invention is to provide a kind of impulsive noise that can detect in the digital audio and video signals, and recover the method for original audio signal.

In order to achieve the above object, technical scheme of the present invention has provided a kind of method that detects and eliminate impulsive noise in the digital audio and video signals,

Step 1, the sound signal that contains impulsive noise is sampled;

Step 2, importing contain the sampled audio signal data of impulsive noise, it is characterized in that:

Step 3, the signal data sample is divided into length in chronological order is 512 sample data frame, and each sample data frame is carried out following processing respectively:

Step 3.1, in current sample data frame, each the sample x (n) in the sample of signal vector x is tested calculating, obtain amplitude, frequency change degree estimated value F and the pulse number C of current data frame;

Step 3.2, obtain the detection threshold T of pulse position and width according to amplitude, frequency change degree estimated value F and the pulse number C of the current data frame that obtains by step 4 ₂, use threshold value T ₂Calculate pulse conceptual vector p;

Step 3.3, the sample of signal vector x is set up P, 20≤P≤2000, the automatic regression model on rank, obtaining length is the model parameter a of P;

Step 3.4, use a model in current sample data frame parameter a and pulse conceptual vector p carry out interpolation to each sample x (n), obtain repairing sample value

Step 3.5, calculate the average error d of all sample interpolations in the noise spike section _k, repair sample value according to this value correction

, the final reparation sample value x ' that obtains sample x (n) (n);

Step 3.6, repeating step 3.1 to step 3.5 until all sample data frames are all handled one time.

The present invention utilizes the weighted difference vector of the interior adjacent signal samples of certain neighborhood in the signal, amplitude, frequency change degree estimated value according to this segment signal detect pulse number, determine the noise spike detection threshold, the envelope that re-uses based on sample of signal weighted difference vector comes the width of estimating noise pulse, set up the autoregressive model of signal, according to the correlativity of sound signal, use has been repaired or unpolluted adjacent signal samples reparation is polluted signal segment by impulsive noise, recovers original signal.The advantage of this method is to make a distinction jump signal in the sound signal and noise spike, the width of estimating noise pulse, and can revise the interpolation error that produces in the repair process.

Advantage of the present invention is to make a distinction changing unexpected normal signal and noise spike, and can calculate the width of pulse, simultaneously, because the envelope that adopts signal weighting difference mean vector to form comes detection noise pulse position and width, thereby can handle single isolated, a plurality of continuous or a plurality of overlapping noise spikes.

Description of drawings

The process flow diagram that Fig. 1 handles the sample data frame data for the present invention;

Fig. 2 changes the valuation synoptic diagram for the signal amplitude-frequency;

Fig. 3 is band noise signal figure;

Fig. 4 is weighted difference and weighted difference mean value signal figure;

Fig. 5 is pulse position and width signal figure;

Signal graph after the method that Fig. 6 passes through through the present invention for the signal of Fig. 3 is handled.

Embodiment

Specify the present invention below in conjunction with embodiment.

Embodiment

As described in Figure 1, for a kind of process flow diagram that detects and eliminate the method for impulsive noise in the digital audio and video signals provided by the invention, the steps include:

Step 1, the sound signal that contains impulsive noise is sampled;

Step 2, importing contain the sampled audio signal data of impulsive noise;

Step 3, the signal data sample is divided into length in chronological order is 512 sample data frame, as shown in Figure 3, each sample data frame is carried out following processing respectively:

Step 3.1, in current sample data frame, each the sample x (n) in the sample of signal vector x is tested calculating, obtain amplitude, frequency change degree estimated value F and the pulse number C of current data frame;

Step 3.1.1, to each the sample x (n) in the sample of signal vector x in the sample data frame, calculate its 2-neighborhood weighted difference score value s (n) by following formula:

$s\left(n\right)=\frac{|x(n-2)-4x(n-1)+6x\left(n\right)-4x(n+1)+x(n+2)|}{T},$ Wherein, T is the signals sampling interval;

The average value mu of step 3.1.2, calculating 2-neighborhood weighted difference score value s (n) _s:

${\mathrm{\μ}}_s=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}s\left(n\right);$

The standard deviation sigma of step 3.1.3, calculating 2-neighborhood weighted difference score value s (n) _s:

${\mathrm{\σ}}_s=\sqrt{\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{[s\left(n\right)-{\mathrm{\μ}}_s]}^2};$

Step 3.1.4, determine pulse detection threshold value T ₁: T ₁=σ _s* K ₁, wherein, K ₁Be an adjustable constant, by the order of severity setting of user according to signal noise, its span is 1～10 when reality is used;

Each sample x (n) among step 3.1.5, the use threshold value T1 detection signal sample vector x, obtaining length is the pulse mark vector p of N, n the element p (n) among the p is:

$p\left(n\right)={\&Integral;}_0^1\begin{array}{cc}\mathrm{if}& s\left(n\right)>={\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="A200910045967E00123.png"\; state="\{\{state\}\}">T</figure-callout>}}_1\\ \mathrm{if}& s\left(n\right)<{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="A200910045967E00123.png"\; state="\{\{state\}\}">T</figure-callout>}}_1\end{array},$ Then umber of pulse C is that element value is the number of 1 fragment continuously among the vectorial p;

The amplitude of step 3.1.6, current data frame, frequency change degree estimated value F calculate by following formula:

$F=\underset{i=-3}{\overset{3}{\mathrm{\&Sigma;}}}f\left(i\right),$ Wherein, the computing method of f (i) can illustrate that in Fig. 2, the time shaft of t for representing with sample, y are signal amplitude value in the current data frame by synoptic diagram 2, curve representation signal x, and 7 horizontal linears are represented y=σ _xI (i is-3～3 integer), f (i) are signal x and straight line bunch y=σ in the sample data frame _xThe number of times sum that i intersects, σ _xStandard deviation for signal x;

Step 3.2, obtain the detection threshold T of pulse position and width according to amplitude, frequency change degree estimated value F and the pulse number C of the current data frame that obtains by step 4 ₂, use threshold value T ₂Calculate pulse conceptual vector p;

Step 3.2.1, be identified for carrying out the detection threshold T of noise spike position and width detection ₂: T ₂=σ _s* K ₂, wherein, σ _sBe the standard deviation of trying to achieve, K by step 3.1.3 ₂Be an adjustable constant when reality is used, its span is 1～10, and concrete value determines that by the umber of pulse C that tries to achieve by step 3.1.5 and the amplitude of trying to achieve by step 3.1.6, frequency change degree estimated value F table one has provided parameter K ₂Computing method, wherein, H is the sample frequency of sound signal, symbol " * " is represented any value;

Table one

K 2	F/H×10 5	C
			10	>177	＊
10	73～177	>9
			3	73～177	3～9
2	73～177	<3
			3	<73	>9
2	<73	3～9
			1	<73	<3

The same length mean vector v of the described 2-neighborhood of step 3.2.2, calculation procedure 3.1.1 weighted difference score value s (n), its n element v (n) is determined by following formula:

$v\left(n\right)=\frac{2}{3}\underset{m=n-1}{\overset{n+1}{\mathrm{\&Sigma;}}}s\left(m\right);$ Fig. 4 is that the noise spike of sample data frame shown in Figure 3 detects synoptic diagram, wherein, and level point line expression threshold value T ₂, last outside envelope solid line is represented the mean vector v of signal weighting difference vector, the dotted line below the envelope solid line is represented signal weighting difference vector s.

Step 3.2.2, use noise spike detection threshold T ₂Recomputate the described pulse mark vector of step 3.1.5 p, its n element value determined by following formula:

$p\left(n\right)={\&Integral;}_0^1\begin{array}{cc}\mathrm{if}& v\left(n\right)>={\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="A200910045967E00123.png"\; state="\{\{state\}\}">T</figure-callout>}}_2\\ \mathrm{if}& v\left(n\right)<{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="A200910045967E00123.png"\; state="\{\{state\}\}">T</figure-callout>}}_2\end{array},$ In the pulse mark vector,, represent that then interior the sample of sample data frame is in noise spike if the value of element is 1; If the value of element is 0, represent that then sample is the normal audio sample of signal, Fig. 5 is the pulse mark vector figure of sample data frame shown in Figure 3;

Step 3.3, the sample of signal vector x is set up P, 20≤P≤2000, the automatic regression model on rank, obtaining length is the model parameter a of P;

Step 3.4, use a model in current sample data frame parameter a and pulse conceptual vector p carry out interpolation to each sample x (n), obtain repairing sample value

:

$\left\{\begin{array}{ccc}\hat{x}\left(n\right)=\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}a\left(i\right)x(n-i)& \mathrm{if}& p\left(n\right)=1\\ \hat{x}\left(n\right)=x\left(n\right)& \mathrm{if}& p\left(n\right)=0\end{array}\right.;$

Step 3.5, calculate the average error dk of all sample interpolations in the noise spike section, repair sample value according to this value correction

The final reparation sample value x ' that obtains sample x (n) (n);

The even distribution this section in of interpolation error in step 3.5.1, the noise segment of supposition, then the average error d of all sample interpolations in k noise spike section _kFor:

$d_k=[x\left(R_{k-\mathrm{end}}\right)-\hat{x}\left(R_{k-\mathrm{end}}\right)]/L_k,$ Wherein, R _K-endBe the sequence number of last sample in interior k the noise spike section of sample data frame, x (R _K-end) be sample R _K-endValue,

Be R _K-endThe reparation sample value of individual sample, L _kThe length of representing k pulse, its solution procedure is: make that k pulse starts from R in the sample data frame _K-start, end at R _K-end, then the length of k pulse is L _k=R _K-end-R _K-start

Step 3.5.2, use average error d _kTo the reparation sample value in the noise segment

Revise, n the sample interpolation that drops on k noise segment in the sample data frame be modified to: $x\&prime;\left(n\right)=\hat{x}\left(n\right)+d_k,$ The reparation sample value x ' that finally obtains former sample x (n) (n), Fig. 6 is the reparation figure of sample of signal frame shown in Figure 3.

Step 3.6, repeating step 3.1 to step 3.5 until all sample data frames are all handled one time.

After band noise signal shown in Figure 3 handled through method provided by the invention, obtain signal shown in Figure 6, by more as can be known, method provided by the invention can make a distinction unexpected normal signal and the noise spike of variation.

Claims (5)

1. method that detects and eliminate impulsive noise in the digital audio and video signals,

Step 1, the sound signal that contains impulsive noise is sampled;

Step 2, importing contain the sampled audio signal data of impulsive noise, it is characterized in that:

Step 3, the signal data sample is divided into length in chronological order is 512 sample data frame, and each sample data frame is carried out following processing respectively:

Step 3.1, in current sample data frame, each the sample x (n) in the sample of signal vector x is tested calculating, obtain amplitude, frequency change degree estimated value F and the pulse number C of current data frame;

Step 3.2, obtain the detection threshold T of pulse position and width according to amplitude, frequency change degree estimated value F and the pulse number C of the current data frame that obtains by step 4 ₂, use threshold value T ₂Calculate pulse conceptual vector p;

Step 3.3, the sample of signal vector x is set up P, 20≤P≤2000, the automatic regression model on rank, obtaining length is the model parameter a of P;

Step 3.4, use a model in current sample data frame parameter a and pulse conceptual vector p carry out interpolation to each sample x (n), obtain repairing sample value

(n);

Step 3.5, calculate the average error d of all sample interpolations in the noise spike section _k, repair sample value according to this value correction

(n), obtain the final reparation sample value x ＇ (n) of sample x (n);

Step 3.6, repeating step 3.1 to step 3.5 until all sample data frames are all handled one time.

2. a kind of method that detects and eliminate impulsive noise in the digital audio and video signals as claimed in claim 1 is characterized in that described step 3.1 comprises the following steps:

Step 3.1.1, to each the sample x (n) in the sample of signal vector x in the sample data frame, calculate its 2-neighborhood weighted difference score value s (n) by following formula:

$s\left(n\right)=\frac{|x(n-2)-4x(n-1)+6x\left(n\right)-4x(n+1)+x(n+2)|}{T},$ Wherein, T is the signals sampling interval;

The average value mu of step 3.1.2, calculating 2-neighborhood weighted difference score value s (n) _s:

${\mathrm{\&mu;}}_s=\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}s\left(n\right);$

The standard deviation sigma of step 3.1.3, calculating 2-neighborhood weighted difference score value s (n) _s:

${\mathrm{\&sigma;}}_s=\sqrt{\frac{1}{N}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{[s\left(n\right)-{\mathrm{\&mu;}}_s]}^2}$

Step 3.1.4, determine pulse detection threshold value T ₁: T ₁=σ _s* K ₁, wherein, K ₁Be an adjustable constant, by the order of severity setting of user according to signal noise, its span is 1～10 when reality is used;

Step 3.1.5, use threshold value T ₁Each sample x (n) among the detection signal sample vector x, obtaining length is the pulse mark vector p of N, n the element p (n) among the p is:

$p\left(n\right)={\&Integral;}_0^1\begin{array}{c}\mathrm{if\; s}\left(n\right)>=T_1\\ \mathrm{if\; s}\left(n\right)<T_1\end{array},$ Then umber of pulse C is that element value is the number of 1 fragment continuously among the vectorial p;

The amplitude of step 3.1.6, current data frame, frequency change degree estimated value F calculate by following formula:

$F=\underset{i=-3}{\overset{3}{\mathrm{\&Sigma;}}}f\left(i\right),$ Wherein, f (i) passes straight line bunch y=σ for signal curve in the sample data frame _xThe number of times of i, σ _xStandard deviation for the sample of signal vector x.

3. a kind of method that detects and eliminate impulsive noise in the digital audio and video signals as claimed in claim 2 is characterized in that described step 3.2 comprises the following steps:

Step 3.2.1, be identified for carrying out the detection threshold T of noise spike position and width detection ₂: T ₂=σ _s* K ₂, wherein, σ _sBe the standard deviation of trying to achieve, K by step 3.1.3 ₂Be an adjustable constant when reality is used, its span is 1～10, and concrete value is determined by the umber of pulse C that tries to achieve by step 3.1.5 and the amplitude of trying to achieve by step 3.1.6, frequency change degree estimated value F;

The same length mean vector v of the described 2-neighborhood of step 3.2.2, calculation procedure 3.1.1 weighted difference score value s (n), its n element v (n) is determined by following formula:

$v\left(n\right)=\frac{2}{3}\underset{m=n-1}{\overset{n+1}{\mathrm{\&Sigma;}}}s\left(m\right);$

Step 3.2.2, use noise spike detection threshold T ₂Recomputate the described pulse mark vector of step 3.1.5 p, its n element value determined by following formula:

$p\left(n\right)={\&Integral;}_0^1\begin{array}{c}\mathrm{if\; v}\left(n\right)>=T_2\\ \mathrm{if\; v}\left(n\right)<T_2\end{array},$ In the pulse mark vector,, represent that then interior the sample of sample data frame is in noise spike if the value of element is 1; If the value of element is 0, represent that then sample is the normal audio sample of signal.

4. a kind of method that detects and eliminate impulsive noise in the digital audio and video signals as claimed in claim 1 is characterized in that described step 3.4 comprises the following steps:

$\left\{\begin{array}{ccc}\hat{x}\left(n\right)=\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}a\left(i\right)x(n-i)& \mathrm{if}& p\left(n\right)=1\\ \hat{x}\left(n\right)=x\left(n\right)& \mathrm{if}& p\left(n\right)=0\end{array}\right..$

5. a kind of method that detects and eliminate impulsive noise in the digital audio and video signals as claimed in claim 1 is characterized in that described step 3.5 comprises the following steps:

The even distribution this section in of interpolation error in step 3.5.1, the noise segment of supposition, then the average error d of all sample interpolations in k noise spike section _kFor:

$d_k=[x\left(R_{k-\mathrm{end}}\right)-\hat{x}\left(R_{k-\mathrm{end}}\right)]/L_k,$ Wherein, R _K-endBe the sequence number of last sample in interior k the noise spike section of sample data frame, x (R _K-end) be sample R _K-endValue,

Be R _K-endThe reparation sample value of individual sample, L _kThe length of representing k pulse, its solution procedure is: make that k pulse starts from R in the sample data frame _K-start, end at R _K-end, then the length of k pulse is L _k=R _K-end-R _K-start

Step 3.5.2, use average error d _kTo the reparation sample value in the noise segment

Revise, n the sample interpolation that drops on k noise segment in the sample data frame be modified to: $x\&prime;\left(n\right)=\hat{x}\left(n\right)+d_k,$ Finally obtain the reparation sample value x ＇ (n) of former sample x (n).

Images